FaceShield-pre10K, FaceShield-sft45K
收藏资源简介:
FaceShield数据集是一组为面部反欺骗任务量身定制的多模态大型语言模型(MLLM)的预训练和监督微调(SFT)数据集。该数据集由多轮对话和图像组成,涵盖了12种不同的攻击类型,旨在帮助MLLM在面部识别系统中识别欺骗行为。数据集包括预训练数据集FaceShield-pre10K和监督微调数据集FaceShield-sft45K,这两个数据集均由基于预先定义的提示的多轮对话生成。FaceShield数据集旨在解决面部识别系统中的欺骗攻击问题,如打印、重放和3D可穿戴面具等,以提高系统的准确性和可靠性。
The FaceShield dataset is a pre-training and supervised fine-tuning (SFT) dataset suite tailored for facial anti-spoofing tasks for multimodal large language models (MLLMs). Comprising multi-turn dialogues and images, this dataset covers 12 distinct attack types, with the goal of helping MLLMs identify spoofing behaviors in facial recognition systems. The dataset includes two subsets: the pre-training dataset FaceShield-pre10K and the supervised fine-tuning dataset FaceShield-sft45K, both of which are generated via multi-turn dialogues based on pre-defined prompts. The FaceShield dataset is intended to resolve spoofing attack issues in facial recognition systems, such as printing, replay attacks, and 3D wearable masks, to enhance the accuracy and reliability of these systems.

- 1FaceShield: Explainable Face Anti-Spoofing with Multimodal Large Language ModelsShijiazhuang Tiedao University, Shanghai Jiao Tong University, UCLA, GRGBanking, Great Bay University, Macao Polytechnic University, Shenzhen Campus of Sun Yat-sen University · 2025年



